CLI prompt/status line is cut off or splits badges in narrow terminal panes

[yxlyx]

Summary

The interactive CLI prompt/status line is not width-aware. In a narrow terminal pane it wraps or appears cut off in the middle of a badge; for example, the visible line can end at:

[gpt-5.6-sol · High · co█

This makes the cursor appear to be inside the codex provider badge and makes it unclear where user input begins.

Environment

• graff 0.0.201
• macOS / Ghostty (TERM=xterm-ghostty)
• Codex model gpt-5.6-sol

Reproduction

Start graff in a narrow terminal pane, or reproduce under tmux with a 28-column pane:

tmux new-session -d -x 28 -y 12 -s graff-narrow \
  "graff --model gpt-5.6-sol --no-telemetry"
tmux capture-pane -p -t graff-narrow

The status prompt physically wraps like this:

[gpt-5.6-sol · High · codex
· cwd /path/to/project
] ›

Depending on the pane/viewport, only the first partial row is visible, which looks like a truncated prompt with the cursor embedded in the provider label.

Expected

The prompt should adapt to available width. Possible behavior:

• omit lower-priority metadata (cwd, token/cache/cost detail) at narrower widths;
• never split a badge such as codex across the edge;
• and/or put a short input marker on a clean new row.

The cursor/input location should remain visually unambiguous.

Likely cause

Agent.prompt in src/agent.zig emits the full model, effort, provider, modes, cwd, and optional context/cache/cost string without budgeting against terminal width. readline.zig checks the cursor column afterward and may move input to another row, but by then the status string itself has already wrapped.

A narrow-width PTY regression covering the main Agent.prompt (not only the model/settings pickers) would help.

[justrach]

A fix was attempted on fix/209-narrow-prompt-width and rejected in review — not merged into release/v0.0.217. It does fix the reported wrapping (verified over a real PTY at 28→200 cols, no badge split at any width), but it introduces a worse problem and has no coverage.

Major — it silently misreports the session's mode. To fit a narrow pane it drops badges by priority, and the ones it drops include Plan, Strict and Ultracode. Measured against the built binary with all three active: at 36 cols the prompt renders just [deepseek-v4-pro · Medium] › — Plan, Strict, Ultracode and Privacy:Local all gone, with no ellipsis or truncation marker. In this issue's own 28-col tmux repro, a user who has run /strict sees no Strict badge and reasonably concludes they are in normal mode while the agent is enforcing tool-only turns. Before the change those badges wrapped but stayed visible. This issue's Expected section sanctioned dropping cwd and the token/cost detail — not the mode badges — and suggested a short input marker on a clean row as the information-preserving alternative.

Major — the fix is untested. Mutating the budget to std.math.maxInt(usize), which restores the exact pre-fix unbudgeted rendering and re-opens this issue verbatim, still leaves the suite at 366/366 green. The three new tests only exercise a 5-line fitsSegment helper and re-derive the layout arithmetic by hand; none of them calls prompt(). Note a working regression already exists on the sibling release/v0.0.203 lineage — scripts/test-pty-narrow-prompt.py from 129031c, with 28-col and wide-pane assertions, exactly what this issue asked for. It was not ported; instead the two existing PTY tests were widened 120→240 cols, so net width coverage went down.

Minor, same lineage: 129031c also moved dispWidth/codepointWidth into term.zig so the budget used true display columns. Those still exist here at src/repl_markdown.zig:118-153; this attempt used a byte-length proxy, which over-charges CJK/emoji cwd paths by up to 50% and drops cwd from panes where it would fit.

Recommended path: port 129031c rather than re-derive it — it already has the PTY regression and the display-width handling — and keep mode badges non-droppable, truncating cwd/meter first with an explicit marker.

[justrach]

Fixed in v0.0.226 (PR #328).

Merged the width-budgeted prompt line. Two branches existed; took the one that extracts the budgeting into agent_prompt.zig (agent.zig was exactly at the 600-line cap) rather than the one that grows agent.zig.

Tests went 449 -> 575 on this release; every Zig fix carries a pinning test that was falsified independently (revert the production change, confirm the test goes red naming itself, restore).